Neuroimaging the language network with a parsing algorithm John Hale Department of Linguistics and Cognitive Science Program Cornell University Jonathan Brennan Department of Linguistics University of Michigan Wen-Ming Luh MRI Facility and Department of Biomedical Engineering Cornell University Christophe Pallier INSERM-CEA Cognitive Neuroimaging Unit, Neurospin center, Université Paris-Saclay CNL Shohini Bhattasali Department of Linguistics Cornell University Conclusion: activation in frontal & temporal areas correlates with transient workload of bottom-up parser Results: per-subject t-maps S NP NNP Boa NNS constrictors VP VB swallow NP PRP$ their NN prey NN whole PP IN without NP VBG chewing 1 2 1 1 1 2 1 8 word time offset bottom-up parser actions freq per million in movie subtitles boa 17.584 1 1.29 constrictors 18.321912 2 0.16 swallow 18.811869 1 12.73 their 18.984 1 655.16 prey 19.45755 1 5.51 whole 19.913324 2 385.49 without 20.334 1 354.65 chewing 20.9022 8 5.51 Method: parser action count, convolved with HRF enters into regression against observed BOLD signal Type to enter text Question: which parts of the brain work like a parser during naturalistic comprehension? materials: 15K words from literary text four comprehension questions after each section (~ 15 minutes) What does the narrator most often discuss with grown-ups? (a) Boa constrictors, primeval forests, and stars (b) Tiaras, diamonds, gold, and platinum (c) Bridge, golf, politics, and neckties (d) Drawings, paintings, and sculptures multi-echo fMRI •Multi-echo T2*-weighted fMRI images were acquired with echo-planar imaging (EPI) at 3 echo times (TEs) – 12.8, 27.5 and 43 ms – opposed to conventional single- echo EPI (Kundu et al 2012) • •Multi-echo fMRI acquisition allows for differentiating signal changes due to Blood Oxygen Level Dependent (BOLD) effects and non-BOLD artifacts such as head motion and cardiac pulsation based on goodness of fit to BOLD biophysical model • •Independent Component Analysis (ICA) was first applied to multi-echo data and derive spatial components subject to sorting based on statistical scores associated with BOLD and non-BOLD signal changes • •Multi-echo fMRI data were de-noised by removing non-BOLD fluctuations from time series data to reliably and robustly improve data quality Participants: college-age, right handed native English-speakers Results: group analysis N=32 Discussion: peaks observed in left angular gyrus and left inferior frontal gyrus, p<0.05 FWE Activation in superior temporal sulcus visible at slightly lower thresholds trees obtained using Stanford Parser D. Klein and C. D. Manning. Accurate unlexicalized parsing. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 423–430, July 2003. for more on metrics J. Hale Automaton Theories of Human Sentence Comprehenson. CSLI Publications 2014.